ทันเอไอ

Researchers Propose New Framework Defining 'Identity' of AI Personality Clones, Measured Externally Instead of Asking About Consciousness

An arXiv paper proposes a way to measure how 'authentic' an AI that mimics a person really is, through external observation — sidestepping the big question of consciousness entirely

As AIs built to mimic the personalities of real people (personality clones) move closer to everyday life, a pressing question follows: how do we know whether a given AI truly 'is' that person? A new paper published on arXiv, titled 'Identity from the Outside: A Conceptual Framework and Research Program for AI Personality Clones,' answers this with a different approach — measuring identity from the 'outside,' through observable behavior, while removing the philosophical problem of consciousness from the equation altogether.

The authors note that the word 'identity' is commonly used with three conflated meanings, so the paper separates them into three distinct criteria: fidelity to the original person, human-likeness in general, and individuality (being unlike anyone else). A given AI may pass some of these criteria while failing others.

At the heart of the work is breaking 'observable identity' down into six factors: substrate, dispositions, memory, update dynamics, context, and exogenous contingencies. The paper also formulates a measure of 'indiscernibility' from an observer's point of view and proposes evaluating the contribution of each factor through randomized ablation experiments.

A key proposal is that long-term identity should be measured using 'climate fidelity' — the consistency of the probabilities with which the AI responds in various ways — rather than trying to force the AI to retrace an identical conversational path, like replaying a recording. The authors also predict that if an AI knows (or is designed to know) that it can be versioned and modified (versionability), this is likely to erode the credibility of its long-term identity.

That said, it must be stressed that this is a conceptual framework and a future research program — not empirical experimental results — and it is a preprint on arXiv that has not yet undergone peer review, so it should be read with the usual caution.